Automatic Classification of Woven Fabric Structure Based on Computer Vision Techniques

Size: px

Start display at page:

Download "Automatic Classification of Woven Fabric Structure Based on Computer Vision Techniques"

Shanon Brown
6 years ago
Views:

1 Journal of Fiber Bioengineering and Informatics 8:1 (215) doi:1.3993/jfbi32157 Automatic Classification of Woven Fabric Structure Based on Computer Vision Techniques Xuejuan Kang a,, Mengmeng Xu b, Junfeng Jing b a Electrical Engineering Department, Xi an Aeronautical University, Xi an 7177, China b School of Electronics and Information, Xi an Polytechnic University, Xi an 7148, China Abstract Traditionally woven fabric structure classification is based on manual work in textile industry. This paper proposes an automatic approach to classify the three woven fabrics: plain, twill and satin weave. Firstly 2-D wavelet transform is used to obtain low frequency sub-image in order to reduce the analysis of fabric images. Then graylevel co-occurrence matrix (GLCM) and Gabor wavelet are adopted to extract the texture features of pre-processing fabric images. Finally Probabilistic Neural Network (PNN) is applied to classify the three basic woven fabrics. The experimental results demonstrate that the proposed method can automatically, efficiently classify woven fabrics and obtain accurate classification results (93.33%). Keywords: Woven Fabric Structure; Automatic Classification; 2-D Wavelet Transform; GLCM; Gabor Wavelet; PNN 1 Introduction In traditional textile industry, analysis and recognition of woven fabric structure mostly depend on manual inspection, which requires a long time and many professional workers. To improve work efficiency, it is necessary to propose an innovative and efficient method for fabric structure recognition. With the development of technology, the application of computer vision and image processing is becoming more dominant. Image processing has been introduced into texture classification, which can automatically and accurately classify woven fabric structure [1]. The analysis of fabric texture [2] has been studied since the mid-198s in Japan, which progresses from optical computing to digital image processing. Recently, some relevant researches have been developed for automatic analysis of woven fabric structure. Haralick et al. [3] proposed a method of image graylevel co-occurrence matrix, and took its four feature parameters of energy, contrast, correlation, entropy as image features to identify fabric images. Melendez et al. [4] used The authors would like to thank for the support of Scientific Research Program Funded by Natural Science Foundation of China (No ); Shaanxi Provincial Education Department (No. 213JK184); and Shaanxi Science and Technology Research and Development Project (No. 213K7-32). Corresponding author. address: @qq.com (Xuejuan Kang) / Copyright 215 Binary Information Press & Textile Bioengineering and Informatics Society March 215

2 7 X. Kang et al. / Journal of Fiber Bioengineering and Informatics 8:1 (215) the convolution images by Gabor filters as texture features. These feature extraction methods have good recognition results. Hu [5] put forward a method of fabric automatic classification based on Bayesian statistics. Shin et al. [6] developed an unsupervised recognition method using fuzzy c-means clustering in the spatial domain. Salem et al. [7] used support vector machine to classify fabric weave patterns. Another identification method [8, 9] analyzed warp and weft floats to determine the fabric weave patterns. These methods can successfully classify several woven fabrics. However, real-time and fault-tolerant abilities of current woven fabric classification methods are low. Hence, this paper introduces an approach for recognition and classification of woven fabrics with real-time and fault-tolerant abilities based on computer vision technique. Firstly, pre-processing image is decomposed into 7 sub-images by two layer wavelet transform, and low frequency subimage LL2 are taken as processing sample to reduce the analysis of fabric images. Then, GLCM and Gabor wavelet are used to extract texture features of woven fabrics. Finally, an appropriate classifier, probabilistic neural network, is applied to recognize woven fabrics in the classification phase. The flow chart of fabric texture image recognition is shown as Fig. 1, and it illustrates the recognition process. Fabric image acquisition and pre-processing 2D wavelet transform Texture features calculation GLCM Gabor wavelet Probabilistic neutral network Recognition of woven fabric Fig. 1: Flow chart of fabric texture image classification 2 Experimental Methods D Wavelet Transform In order to reduce the amount of calculation and remove noise, 2-D wavelet transform [1] is adopted to process woven fabric images. Wavelet transformation is a local transformation between time domain and frequency domain, which can extract information from the signal effectively. It can conduct multi-scale detail analysis for function and signal by flex and translation, and solve

3 X. Kang et al. / Journal of Fiber Bioengineering and Informatics 8:1 (215) the problem that Fourier transforms cannot solve. 2-D wavelet transform is made by sampling interval in rows and columns after inner product between an original image and a wavelet basis image. Each convolution can decompose the 1-D convolution on rows and columns, because scale function and wavelet function are separable. After the first layer of wavelet transform, original image is decomposed into four sub-images: a lowfrequency approximate sub-image LL1 and three high-frequency detailed sub-images (a horizontal sub-image HL1, a vertical sub-image LH1, and a diagonal sub-image HH1). In the second layer of wavelet decomposition, the low-frequency part (LL1) is only decomposed as mentioned above, it then produces the different frequency band outputs. Fig. 2 shows the diagram of two layer wavelet decomposition. f(x, y) the original image LL1 LH1 HL1 HH1 LL2 HL2 LH2 HH2 LH1 HL1 HH1 (a) (b) (c) Fig. 2: The two layer wavelet decomposition. (a) The original image. (b) The first layer wavelet decomposition. (c) The second layer wavelet decomposition 2.2 Graylevel Co-occurrence Matrix The Graylevel Co-occurrence Matrix (GLCM) of an image can reveal comprehensive information of gray level [11, 12], which are about directions, adjacent interval and rangeability. So far, it has been proved to be a good method of texture analysis in theory and experiment. GLCM has been brought out in 1973 by Haralick [3]. It describes grey value i and j, the two pixels all appeared probability matrix C(i, j, d, θ) in θ direction and d distance. θ is the position angle between two pixels. Distance d is the distance between two pixels, generally according to the test image to determine. Generally, four common statistics are used to extract image texture characteristics: angular second moment (energy), contrast, correlation and entropy, which are not only convenient for computation but also giving a higher accuracy of classification. Angular second moment (Energy): ASM = M i=1 N C(i, j, d, θ) 2 (1) j=1 Contrast: Correlation: CON = COR = M i=1 M i=1 N (i j) 2 C(i, j, d, θ) (2) j=1 N j=1 ijc(i, j, d, θ) µ x µ y σ x σ y (3)

4 72 X. Kang et al. / Journal of Fiber Bioengineering and Informatics 8:1 (215) with µ x = M N ic(i, j, d, θ), µ y = i=1 j=1 M N jc(i, j, d, θ) i=1 j=1 σ x = M N (i µ x ) 2 C(i, j, d, θ), σ y = i=1 j=1 M N (j µ y ) 2 C(i, j, d, θ) i=1 j=1 Entropy: M N ENT = C(i, j, d, θ)lgc(i, j, d, θ) (4) i=1 j=1 It is defined that the value of direction θ of, 45, 9, 135, and distance d is 1 in GLCM method. Therefore, 16-dimensional GLCM features of angular second moment, contrast, correlation and entropy in four directions and one distance are obtained for each identified fabric image. 2.3 Gabor Filter As the good multi-scale and multi-orientation decomposition of 2-D Gabor filter in frequency domain, it is applied into computer vision and image processing widely. 2-D Gabor function is the result of modulation by 2-D Gaussian function [13] with trigonometric function. By using Gabor function, the local frequency domain information in different scales and different orientations can be obtained. A 2-D Gabor function g(x, y) and its Fourier transform G(µ, ν) can be expressed as: ( 1 g(x, y) = 2πσ x σ y ) exp G(µ, ν) = exp { 1 2 { 1 2 ( x 2 1 σ 2 x + y2 1 σ 2 y ( (µ f ) 2 σ 2 µ )} exp(2πjf x 1 ) (5) )} + ν2 σν 2 where, f is mid-frequency, x 1 = (xcosθ+ysinθ), y 1 = ( xsinθ+ycosθ), θ = nπ and K is the total K number of orientations, n [, K]. σ x, σ y are the variances along x axis and y axis respectively. σ µ = 1 2πσ x, σ ν = 1 2πσ y. The real part of 2-D Gabor function shown as Eq. (7) acts as an even symmetric Gabor filter to make image smoothing. While the imaginary part of 2-D Gabor function is used for detecting fabric edge part as an odd symmetric filter to be shown as Eq. (8). The relationship of two portions and integrated Gabor filter can be described as Eq. (9). General responses of real part and imaginary part from 2-D Gabor function are exhibited in Fig. 3. ( 1 g e (x, y) = 2πσ x σ y ( 1 g o (x, y) = 2πσ x σ y ) exp { 1 2 ( x 2 1 σ 2 x + y2 1 σ 2 y (6) )} cos(2πf x 1 ) (7) ) { exp 1 ( )} x y2 1 sin(2πf 2 σx 2 σy 2 x 1 ) (8)

X. Kang et al. / Journal of Fiber Bioengineering and Informatics 8:1 (215) 69 79 73 1. 1..5.5.5.5 1. 5 5 5 (a) 5 1.

3: The real part (a) and imaginary part (b) of a basic Gabor function g(x, y) = g e (x, y) + jg o (x, y) (9) Based

little to extract texture features, the real part of 2-D Gabor function is used to extract the texture features.

In order to cover the frequency space of the texture images as best as possible, a reasonable oscillation

4 shows 5 8 Gabor filters. In each scale, filters are spread equally in the 18.

After designing a group of Gabor filters, the filtered images are from the convolution between input image and

Finally, the mean and standard deviation of filter images can be calculated as texture features.

supervised neural network based on Bayesian minimum risk criteria (the Bayesian decision theory).

taking Parzon window function as an activation function.

classification and faster training speed.

5 X. Kang et al. / Journal of Fiber Bioengineering and Informatics 8:1 (215) (a) (b) 5 Fig. 3: The real part (a) and imaginary part (b) of a basic Gabor function g(x, y) = g e (x, y) + jg o (x, y) (9) Based on 2-D Gabor function, because the imaginary part not only requires a mass of calculations, but also contributes little to extract texture features, the real part of 2-D Gabor function is used to extract the texture features. Then Gabor function can be formulated as Eq. (7). In order to cover the frequency space of the texture images as best as possible, a reasonable oscillation frequency and filter orientation must be selected, which can extract the features well. Fig. 4 shows 5 8 Gabor filters. In each scale, filters are spread equally in the 18. In the different scales, Gabor filters play a similar role of microscopes. After designing a group of Gabor filters, the filtered images are from the convolution between input image and each Gabor filters. Finally, the mean and standard deviation of filter images can be calculated as texture features. Therefore, after using Gabor wavelet method, 8-dimensional Gabor features can be obtained for each detected fabric image. 2.4 Probabilistic Neural Network Probabilistic Neural Network [14] is proposed by Specht in 199 and a kind of supervised neural network based on Bayesian minimum risk criteria (the Bayesian decision theory). Besides, it is not only an artificial neural network based on statistical principle but also a feedforward network taking Parzon window function as an activation function. Compared with the traditional neural network, PNN has more significant advantages especially in pattern classification and faster training speed. The PNN consists of four layers (input layer, pattern layer, summation layer and contest layer), and its basic structure is shown in Fig. 5. Input layer passes texture features of training samples to network by linear transfer function. Pattern layer calculates the matching relation between input texture features and each training pattern, and the number of neurons in pattern layer is equal to the sum of each category of training samples, the output of each pattern unit is through following non-linear operator. [ f(x, W i ) = exp (X W ] i) T (X W i ) (1) 2δ 2 where, W i is the linked weight from input layer to pattern layer; δ is the smoothing factor which plays a vital role in the process of classification; X = [x 1, x 2,..., x n ]. In view of this, PNN classifier

4: A group of Gabor filters from five scales and eight orientations x 1 y 1 x

5: The structure of PNN completes the work of nonlinear learning algorithm

Summation layer accumulates estimate probability density which belongs to a

This unit simply adds output which belongs to their own class of pattern

The last layer is contest layer, which is composed of simple threshold

6 74 X. Kang et al. / Journal of Fiber Bioengineering and Informatics 8:1 (215) Fig. 4: A group of Gabor filters from five scales and eight orientations x 1 y 1 x 2 y 2 x n y m Input layer Summation layer Contest layer Pattern layer Fig. 5: The structure of PNN completes the work of nonlinear learning algorithm with high accuracy character using linear learning algorithm. Summation layer accumulates estimate probability density which belongs to a certain class according to Eq. (1). This unit simply adds output which belongs to their own class of pattern layer unit. The last layer is contest layer, which is composed of simple threshold discriminator whose role is to choose a neuron with a maximum posterior probability density in each type as output of the whole system. Neuron with the biggest posterior probability density outputs 1 which corresponds to this class is the pattern category of classified sample, and other

Compared with other methods, PNN classifier does not need to be multiple fully calculated,

In the training pattern samples under certain circumstances, only the smoothing factor

3 Results and Discussion In this paper, experimental samples include plain, twill, and

There are 45 groups woven fabric images which contain 15 groups of plain, twill and satin

Meanwhile, 1 images per class were chosen as the training samples and the rest as testing

7 X. Kang et al. / Journal of Fiber Bioengineering and Informatics 8:1 (215) neurons output. Compared with other methods, PNN classifier does not need to be multiple fully calculated, and can be stably converged to Bayesian optimal solution. In the training pattern samples under certain circumstances, only the smoothing factor needs to be adjusted, PNN network can converge with faster speed. 3 Results and Discussion In this paper, experimental samples include plain, twill, and satin weaves are acquired by Canon- Scan 9F under the same external condition. There are 45 groups woven fabric images which contain 15 groups of plain, twill and satin weaves. Meanwhile, 1 images per class were chosen as the training samples and the rest as testing samples. Fig. 6 shows some samples of experimental Plain Satin Twill Fig. 6: Woven fabric samples (a) 1 2 (b) (c) Fig. 7: Pre-processing effect of fabric image. (a) The histogram of gray fabric image. (b) The histogram of the enhanced fabric image. (c) The gray fabric image. (d) The enhanced fabric image (d)

Category 76 X. Kang et al. / Journal of Fiber Bioengineering and Informatics 8:1 (215) 69 79 woven fabric images.

Furthermore, due to the pixel of gray images that are too concentrated, double-peak Gaussian function for histogram equalization is used to distribute images gray level in lower and higher graylevel

8 Category 76 X. Kang et al. / Journal of Fiber Bioengineering and Informatics 8:1 (215) woven fabric images. The captured RGB images are resized into pixels and converted into gray images for improving processing speed. Furthermore, due to the pixel of gray images that are too concentrated, double-peak Gaussian function for histogram equalization is used to distribute images gray level in lower and higher graylevel so that images contrast can be enhanced. The results of twill woven fabric are shown in Fig. 7. After image pre-processing, 2-D wavelet transform decomposition is used to reduce the size of the original image. Take the plain weave for example, Fig. 8 (b) shows that the low-frequency sub image LL1 with the size of after the first wavelet transform. The LL1 image is then decomposed by the second wavelet transform, and low-frequency sub image LL2 (Fig. 8 (c)) is obtained (a) (b) (c) Fig. 8: The result of two layer wavelet decomposition of the plain weave. (a) The plain weave. (b) The LL1 image. (c) The LL2 image Through 2-D wavelet transform decomposition, the size of fabric images is reduced. Then, 96-dimensional features which include 16-dimensional GLCM features and 8-dimensional Gabor features for each fabric image can be obtained quickly. At the end of the process, 45 woven fabrics can be classified by PNN. In this phase, the particularly important thing is to choose the appropriate value of the smoothing factor δ, which directly affects the final classification results. If δ value is too small, PNN classifier can be used as the most adjacent classifier, which leads to classification results that are not obvious. If δ value is too large, it can cause difficulties in the Training sample number Fig. 9: The training classification results of PNN network with GLCM and Gabor features

9 Category Category X. Kang et al. / Journal of Fiber Bioengineering and Informatics 8:1 (215) process of network computing. Therefore, various δ values are used to train the PNN to obtain optimal classification effect in simulation process. It should be noted that the blue circles and the red stars in Fig. 9 represent training results of sample categories and actual sample categories respectively. Besides, the blue triangles and the red stars in Fig symbolize prediction results of sample categories and actual sample categories. Therefore, if one symbol interlaces another, it means that the samples are classified correctly and vice versa. From Fig. 9, the training effects of 3 training samples achieve the best performance, and training accuracy reached 1%. Besides, network training time is.5 second. For further examining the extrapolation performance, the other 15 testing samples are used to make prediction through applying the successful trained PNN. The testing classification result is shown in Fig. 1. From Fig. 1, it can be easily found that only one test sample is identified as wrong and classification accuracy reaches 93.33%. Comparatively speaking, classification accuracy only using the GLCM features is 86.67% (Fig. 11), and classification accuracy only using the Gabor wavelet features is 8% (Fig. 12) Testing sample number Fig. 1: The testing classification results of PNN network with GLCM and Gabor features Testing sample number Fig. 11: The testing classification results of PNN network only with GLCM features

10 Category 78 X. Kang et al. / Journal of Fiber Bioengineering and Informatics 8:1 (215) Testing sample number Fig. 12: The testing classification results of PNN network only with Gabor features 4 Conclusion An automatic and efficient classification method is proposed in this paper including image preprocessing, feature extraction, fabric pattern recognition and so on. Using 2-D wavelet transform to process fabric image can not only reduce the size of fabric image but also shorten the running time of texture analysis. Then, GLCM and Gabor wavelet are used to extract texture features of fabric images. Compared with BP network and LVQ network, PNN classifier has simple network learning process, faster training speed and a powerful pattern classification ability. PNN is a method of pattern recognition, which has good generalization ability and classification capacity. Furthermore, PNN will not get into local optimal point problem. The experimental results show that the methods of feature extraction and image pattern recognition are accurate and efficient, and woven fabrics recognition system gains the best classification results (93.33%) with a faster speed. With the rapid development of computer technology and image processing technology, woven fabric image automatic identification and classification could promote the development of textile industry. References [1] A. Baykasoglu, L. Özbakir, S. Kulluk, Classifying defect factors in fabric production via DIFACONN-miner: A case study, Expert Systems with Applications, 38(9), 211, [2] T. J. Kang, C. H. Kim, K. W. Oh, Automatic recognition of fabric weave patterns by digital image analysis, Textile Research Journal, 69(2), 1999, [3] R. M. Haralick, K. Shanmugam, I. H. Dinstein, Textural features for image classification, IEEE Transactions Systems, Man and Cybernetics, 3(6), 1973, [4] J. Melendez, M. A. Garcia, D. Puig, Efficient distance-based per-pixel texture classification with Gabor wavelet filters, Pattern Analysis and Application, 11(3-4), 28, [5] J. L. Hu, Textile classification based on Bayesian approach, Journal of Textile Research, 25(1), 24, 48-49

11 X. Kang et al. / Journal of Fiber Bioengineering and Informatics 8:1 (215) [6] C. Y. Shih, J. Y. Lee, Automatic recognition of fabric weave patterns by a fuzzy C-means clustering method, Textile Research Journal, 74(2), 24, [7] Y. B. Salem, S. Nasri, Automatic recognition of woven fabrics based on texture and using SVM, Signal, Image and Video Processing, 4(4), 21, [8] X. Wang, N. D. Georganas, E. M. Petriu, Fabric texture analysis using computer vision techniques, IEEE Transactions on Instrumentation and Measurement, 6(1), 211, [9] R. Pan, W. Gao, J. Liu, H. Wang, Automatic recognition of woven fabric pattern based on image processing and BP neutral network, Journal of the Textile Institute, 12(1), 211, 19-3 [1] J. L. Liu, B. Q. Zuo, Application of wavelet coefficients on similarity description of fabric structural texture, Computer Engineering and Applications, 45, 29, [11] M. Hanmandlu, D. Sujata, D. K. Choudhury, Fabric image defect defection by using GLCM and ROSETTA, International Journal Of Computer Science and Applications, 2(1), 29, 47-5 [12] J. F. Jing, J. Wang, P. F. Li, Y. Li, Automatic classification of woven fabric structure by using learning vector quantization, Procedia Engineering, 15, 211, [13] G. M. Haley, B. S. Manjunath, Rotation invariant texture classification using modified Gabor filters, IEEE Proc. Int. Conf. Image, 1, 1995, [14] D. F. Specht, Probabilistic neutral networks, Neural Networks, 3(1), 199,

Study on fabric density identification based on binary feature matrix

153 Study on fabric density identification based on binary feature matrix Xiuchen Wang 1,2 Xiaojiu Li 2 Zhe Liu 1 1 Zhongyuan University of Technology Zhengzhou, China 2Tianjin Polytechnic University Tianjin,